Scoring

Current scoring for code tasks uses weighted dimensions:

• Test passing: 40%
• Execution success: 25%
• Syntax validity: 15%
• Completeness (LLM judge): 20%

Notes

• Dedicated scorer is implemented for code tasks.
• Other task types currently fall back to code scorer behavior.